One fine-tune stripped a Chinese AI's propaganda from 22% to under 1%

Nous Research released Arete-27B on November 6: Alibaba's Qwen3.8-27B post-trained to reinforce Western values with nothing but supervised fine-tuning — about 53 million tokens, no reinforcement learning, no architecture changes. On Nous's own 803-prompt evaluation, answers adopting the Chinese government's official line fell from 22 percent to 0.9 percent, while factual accuracy rose from 2.32 to 2.98 out of 5 and general capability barely moved. Weights, eval and training data are all public under Apache 2.0.

The lab's claim is that the bias layer is shallow — 'applied late, applied thin' — and cheap to remove, which cuts both ways: the same fine-tune could install any viewpoint. The numbers are self-reported and judged by xAI's Grok 4.7, with no independent rerun yet, and Nous dropped Nvidia's Nemotron 3 Ultra as a teacher over suspected China-linked contamination in its training data.